Family offices hold some of the most sensitive information that exists about a family: net worth, holdings, entity structures, account details, health, security arrangements, philanthropy, and personal matters. AI tools are powerful, but they are also a new place where that information can leak, persist, or be exposed. These are practical rules for handling confidential information when AI is in the loop.
The governing instinct: share the least, in the safest place, and assume anything you paste could persist.
Treat the following as confidential by default. The more of these a piece of information touches, the more carefully it must be handled.
- Names of the family, principals, household members, and staff
- Net worth, balances, holdings, performance, and cash positions
- Entity names, structures, ownership, and governance details
- Account numbers, custodian details, and any financial identifiers
- Trust, estate, and beneficiary information
- Legal matters, disputes, and settlement terms
- Health, medical, and personal information
- Travel plans, residences, and physical-security details
- Counterparty, vendor, and deal information shared in confidence
- Passwords, API keys, tokens, and any credential
Do not put any of the following into a public chatbot, a free consumer AI tool, a browser extension of unknown provenance, or any environment without enterprise data-handling terms:
- Real names of the family, principals, or staff tied to financial detail
- Account numbers, balances, or holdings
- Entity ownership structures and governance specifics
- Trust and estate documents or terms
- Tax returns, financial statements, or capital account detail
- Legal documents, settlement terms, or matters under privilege
- Health or personal information
- Travel itineraries, residence addresses, or security arrangements
- Any credential, key, password, or token
- Counterparty information shared under an NDA or expectation of confidence
If you would not email it to a stranger, do not paste it into a tool whose data terms you have not verified.
- Prefer an enterprise AI environment with written terms covering confidentiality, data retention, and a commitment not to train models on your inputs.
- Verify the data terms before first use, not after. If you cannot confirm how a tool handles your data, treat it as public.
- Avoid unmanaged browser extensions and plug-ins that route your data through third parties.
- Keep a short list of approved tools and route all confidential work through them.
Even in an approved environment, share only what the task requires.
- Redact identifiers the task does not need. A document digest rarely needs the family name; an account number is almost never required for analysis.
- Generalize. Replace "the Smith Family 2019 Irrevocable Trust" with "the Trust" when the specific name adds nothing.
- Use placeholders like
[Principal],[Entity A],[Vendor],[Counterparty]and keep the mapping outside the AI tool. - Split sensitive context from the working text when you can analyze the structure without the identifying detail.
- AI output is working material. Store it under the same access controls as the source documents it was built from.
- Do not forward AI output containing sensitive detail through channels you would not use for the underlying documents.
- Delete working drafts that are no longer needed, consistent with your recordkeeping obligations.
- Persistence. Assume inputs may be retained somewhere. The safest assumption is that a paste is not fully reversible.
- Aggregation. Individually harmless details can become sensitive in combination (a residence + a travel date + a name).
- Over-trust. Fluent output invites you to relax. Sensitivity discipline must not relax with it.
- Convenience drift. The riskiest moment is a busy day when someone reaches for the fastest tool instead of the approved one.
Ask, every time:
- Is this an approved environment with acceptable data terms?
- Have I removed identifiers the task does not need?
- Would I be comfortable if this exact text were exposed?
- Is there a credential or key hiding in what I am about to share? (Never share it.)
- Where will the output live, and is that location appropriately controlled?
If any answer is uncertain, stop and resolve it before sharing.
If confidential information is exposed to the wrong tool or party:
- Stop using that tool for the matter immediately.
- Capture what was shared, when, and where.
- Escalate to the person accountable for the office's risk and, where appropriate, to legal or compliance counsel.
- Review how it happened and adjust the practice so it does not recur.
This document is practical operational guidance, not legal or cybersecurity advice. Confidentiality, privacy, and data-protection obligations vary by jurisdiction and structure; consult qualified advisers for your specific situation.